Information Technology Reference
In-Depth Information
Fig. 4
The graph models of the Bayesian verification process
the object, which is computed using Eq.(9).
p
is the appearance confidence
measure generated by a static people detector (here, we use logistic regression to
convert the output of the detector into a probabilistic interpretation).
p
(
c
|
h
,
o
)
is the
confidence of the hypothesis by the object detector, which in our case is a hypothesis
presence indicator for a certain object class. The final candidates are selected by
thresholding
p
(
h
|
o
)
(
|
,
,
)
o
c
m
h
.
3
Experimental Results
3.1
Data Set
Training set:
We use INRIA image dataset
1
for training, which is totally indepen-
dent to our test video data. The INRIA dataset is both challenging and also does not
contain any motion information. It contains 607 positive training images, together
with their left-right reflections (1214 images in all). A fixed set of 12180 patches are
randomly sampled from 1218 person-free training images. Examples of these im-
ages are shown in Fig. 5. 300 human motion templates are collected from a separate
courtyard sequences, which are used to learn the parameters of the spatial pyramid
Gaussian Mixture. Examples of those are illustrated in Fig. 7.
Test set:
Our test set contains image frames from CCTV video sequences taken from
underground stations and platforms by fixed cameras. One set of images is from a
train platform containing 3710 frames. The other set is from a ticket office area
containing 160 frames. In contrast to other video sequences captured under well-
controlled environment, these sequences present significant lighting changes and
background clutter. Many frames contain multiple people under severe occlusions.
Fig. 6 shows 6 consecutive example frames from the platform scene.
3.2
Detection Results
Examples of detections at the platform area scene from our dynamic detector are
compared and shown against those from a static people detector [3] in Fig. 8. The
detected boxes are displayed on each frame. Fig. 8 (d) shows detection results from
1
The dataset may be downloaded from http://lear.inrialpes.fr/data.
Search WWH ::
Custom Search